Expand description
Compatibility forwarding to the canonical Etalon owner.
Modules§
- adapter
- Prepared operator adapter binding representation digest and residency generation (W3: EOS-031).
- cpu_
lookup - Optimized CPU lookup execution kernel (W6: EOS-060).
- gpu_
schedule - Forge GPU decode/prefill dual-path schedules and occupancy diagnostics (W6: EOS-061).
- moe_
operator - Generic independent and clustered expert execution.
Structs§
- Cluster
Anchor - A cluster anchor containing shared base weights and per-expert deltas (EOS-071).
- Clustered
MoEOperator - Clustered MoE operator holding cluster anchors with shared base projections and per-expert deltas (EOS-071).
- Clustered
MoeScratch - Scratch buffers required for zero-heap clustered MoE evaluation.
- CpuSchedule
Config - Configuration for CPU lookup execution.
- GpuOccupancy
Diagnostics - Analytical occupancy diagnostics for a synthesized Forge schedule.
- GpuOperator
Schedule - Synthesizes a GPU execution schedule descriptor for an operator matrix multiplication.
- Independent
Expert Weights - Independent SwiGLU expert weights (EOS-070).
- LowRank
Delta - A low-rank residual factor delta: delta_W * x = A * (B * x).
- Prepared
Operator - An operator prepared for execution within a residency generation.
- Routed
Expert - Routed expert execution item: (expert_index, gating_weight).
Enums§
- GpuSchedule
Kind - Target execution path on the GPU.
- Prepared
Operator Error - Errors occurring during operator preparation or execution.
Functions§
- dispatch_
clustered_ moe_ step - Execute clustered MoE dispatch step with down-projection pre-accumulation (EOS-071).
- execute_
cpu_ q4k_ lookup - Execute CPU lookup GEMV for GGML Q4_K weights against an input vector.
- precompute_
activation_ luts - Precomputes 16-entry nibble lookup tables for each 32-element subblock in an input vector.